Canvas tools

So on the canvas you'll see a few empty cards appear first, then the images arrive one by one. That isn't stuck: the Agent first fills in the node's name, prompt and model, then submits the generation. Queue and generation progress are drawn on the node, and the timeline adds the result automatically once everything settles (see 12.1).

A card shows the node as it is now — a video that fills in its cover a few minutes later, a node you renamed, and the card follows along. A node gets a card only under the last line that touched it: for the create-then-run pair the card hangs under the "Generate node" line, rather than two identical cards under two lines.

Delete that node later and the card does not disappear with it: it keeps the version it had. The timeline is the record of what this turn did, and a deleted node shouldn't leave a hole in it — the only difference is that clicking the card no longer finds anything on the canvas.

"Not wired" is not a breakage, and it isn't the network either — that action simply isn't available to the Agent yet. It is a settled state, so retrying gets you nowhere; the reply the agent receives also says why, and where the job actually gets done.

Do these yourself, in the interface:

The only nodes it can reach are the ones you reference for it. The least work is to select them on the canvas — selecting is referencing, and a chip turns up in the input box right away (see 12.1).

More than one action spends credits: besides "Generate node", upscale, multi-angle, relight, audio separation, running a whole group and similar canvas tools cost credits too when it runs them for you. Turn off Generate images/videos automatically and it asks you once before every step, showing what that step is about to do. Hit Reject: nothing is charged, and it doesn't quietly retry — it comes back and asks how you want it changed. See 12.6.